Cohere North vs Sema4.ai (2026)

Named integrations reach Google Drive, SharePoint, Slack, Salesforce, Gmail, Outlook and Linear, spanning storage, messaging, CRM, email and project management. The vendor describes these as first-party integrations maintained by Cohere rather than a marketplace of community connectors. The open-ended half is explicit and matters most for an enterprise behind a firewall: North uses the customer's own MCP servers for any custom tool, and Agent Studio documents enhancing agents with custom tools including MCP integration. A bank running North air-gapped needs to reach internal systems no vendor has built a connector for, and MCP is how that happens without waiting on a roadmap. Compass adds a second integration surface through pre-built connectors for retrieval, so systems can serve as knowledge sources as well as action targets.

Breadth across classes is met: SharePoint and SAP are named, with ERP, CRM and data platforms as classes, and the Actions framework reaches any API. Integration is automation as code in Python rather than a connector catalog: the Robocorp libraries reach data, APIs, browsers and desktops, so where no connector exists a customer writes one in a language their engineers already have, with a managed environment supplied. That trades immediate coverage for no ceiling. Tool calling is first class: Actions are decorated Python functions exposed as callable tools, and the same surface serves MCP clients. Zero copy access to the customer's warehouse reaches governed structured data without an integration step. A June 2026 MCP Access Gallery covering Snowflake, Slack, GitHub and Google Workspace has been reported by third parties but is not confirmed on Sema4's own pages.

North Automations, launched around July 2026, is a dedicated orchestration layer. The vendor positions it as moving from isolated task automation to coordinated outcome-driven workflows, with a centralized layer to coordinate diverse agents across an enterprise. Agent-to-agent coordination is named separately on the Agent Studio page. The underlying work is specific: agents automate repetitive tasks, conduct market research, interpret data, and produce documents, tables, charts and slideshows, strengthened by Cohere's acquisition of Ottogrid for market-research automation. Agent-powered workflows are designed around a team's own tools and processes. On control flow, Cohere's announcement says North Automations creates a clear and auditable execution path with loops and branching. Cohere publishes no North product documentation, so how those constructs are built rests on that announcement.

Multi agent execution is specific: agents understand context, reason, take action and collaborate, with fifteen or twenty working together to run entire multi step business processes autonomously. Runbooks are the orchestration primitive. A runbook is written in plain English and describes intent and logic rather than a graph, so the agent reasons about how to satisfy it; business users author them, with the Sai assistant generating drafts and recommending the actions needed, and Studio provides build, test and deploy. Execution underneath is deterministic: Actions are Python, so the reasoning layer plans and a reliable runtime executes. No branching, looping or conditional constructs are named, because runbooks express intent rather than a graph; that suits a different buyer than a graph builder does.

North Automations runs on a schedule: the announcement lists setting the frequency with scheduled runs among its key features, and its example of a daily customer engagement tracker runs each morning and sends a Slack message for review. North is also a workspace where teams chat with agents, integrated with Slack, Gmail, Outlook, Salesforce, Google Drive and SharePoint, so it sits close to where people already work. The gaps: no webhook or external event trigger is named, and because Cohere publishes no North product documentation, the schedule rests on the vendor's announcement.

A schedule creates Work Items for a Worker agent on a standard five field cron cadence in a chosen timezone, with catch up on missed occurrences and last and next run times, set from the schedule picker or through the Schedules API. The Work Item API lets an external system create work items over Bearer authenticated REST, with status callbacks to a webhook. Both wake an agent without a person starting the run. Conversational agents are also reachable in Slack and Microsoft Teams.

Grounding is the product's technical core, served by a purpose-built stack rather than a generic retrieval step. Compass is an intelligent search and discovery system with pre-built connectors, document parsing and a managed index, which is a maintained standing structure. Behind it sit Embed for semantic search and Rerank for retrieval precision across documents, tables and code, so the retrieval pipeline is built from components the same company builds and tunes. Two properties lift this above standard retrieval. Every answer returns citations and transparent reasoning chains, so a reader can verify both the source and the path to the conclusion. And the Command family is optimized for grounded generation rather than adapted to it. A third is structural: because North deploys inside the customer's environment, the index over their proprietary documents is built and held within their own perimeter. For the banks, hospitals and government agencies this product targets, grounding on sensitive material is usually where an AI project stops; here the corpus never leaves. Documented connectors reach Google Drive, SharePoint, Slack and Salesforce among others.

Sema4.ai's documentation describes Knowledge Bases as a semantic layer over unstructured enterprise content such as documents, emails and chat history, built with the Sema4.ai SDK on PostgreSQL with pgvector, with pages on building, maintaining, deploying and querying them; the docs position them as optimized for recall, reasoning and citations. Semantic data models with verified queries give agents a maintained layer over structured databases and files, and Document Intelligence handles document extraction. That is a maintained retrieval structure, which meets the bar.

What Cohere documents is retrieval, not memory. Compass maintains a managed index over the customer's data and North grounds every answer in it with citations, but that is a corpus the customer supplies. Nothing describes an agent retaining conversation context across sessions, carrying a conclusion from one task into the next, or updating what it knows from experience. Agent configuration persists, which is setup rather than accumulated state. The distinction is easy to blur on a product like this because the retrieval is so good: an assistant that always knows your documents can feel like it remembers while holding no state between conversations. Cohere publishes no product documentation for North, so this rests on marketing pages rather than a documentation index. Conversation persistence often ships without being marketed, so this finding could be wrong; a customer-facing user guide would settle it.

What persists is documented, and it is state within a session. Conversational agents accumulate chat threads and worker agents process work items, both tracked in Control Room. A thread is a conversation buffer, and nothing documents an agent reading state written during an earlier session, which holds this at Partial. Runbooks persist, but a runbook is the agent's definition, authored by a person, not state the agent writes and reads back. Third party reporting of a June 2026 release describes persistent memory that retains corrections across runs; it is not confirmed on any Sema4 page.

North Automations names an approval step: the announcement lists identifying the key moments to loop in colleagues for approval as a governance feature. Plan mode lets a person review and edit the approach before an automation is built, and the vendor advises starting with supervised autonomy and setting policies for agent interactions and oversight. Guardrails and granular roles and permissions sit around that, and agents operate within the organization's guardrails. The gaps: the approval step is described in an announcement rather than product documentation, and nothing says what a guardrail actually enforces. Buyers should ask to see where a run pauses for approval.

Sema4.ai's documentation gives Work Items a NEEDS_REVIEW state for when an agent meets a situation that needs human intervention, clarification or a decision, after which a person completes the item or restarts the agent, and its runbook guidance says a good runbook is mostly about when to act and when to escalate, such as flagging new vendors for review. That pause is triggered by the agent following runbook rules, which is agent initiated escalation rather than an approval step the platform imposes before an action, so this is Partial. The Actions framework's is_consequential flag controls whether OpenAI's custom GPT asks a user to approve each action, but that gate belongs to OpenAI's product, not Sema4's. Nothing documented pauses a Sema4 agent before it writes to SAP or moves money; a runtime approval step would take this to Full.

Cohere operates a trust center listing SOC 2 Type II audited annually, with the report available on request under mutual NDA, ISO 27001 for information security management, ISO 42001 for AI management, and the UK government-backed Cyber Essentials standard, plus a documented bug bounty program with published rules of engagement. ISO 42001 is worth naming separately because it certifies the AI management system rather than general information security, and it is unusual. Controls match: granular roles and permissions, agent autonomy policies, full data traceability and audit-ready logs, a multi-jurisdictional Data Processing Addendum, and HIPAA readiness. A name collision to watch: cohere.io is a different company, Caldera Labs Inc. of New York, selling customer-service cobrowsing. It publishes its own security page and its own SOC 2 Type II announcement describing Google Cloud hosting and HITRUST. Those pages do not describe Cohere Inc.; the AI company is cohere.com. A vendor risk aggregator asserts Cohere holds no SOC 2, ISO 27001 or GDPR compliance, which Cohere's own trust center contradicts directly.

The access model is documented: Admin, Builder and Member roles with invitation, domain auto add and account disabling, Builder scoped API keys separate from admin issued service account keys, and organization level SSO with OIDC; the pricing page lists SSO on the Departmental and Enterprise plans. Sema4.ai's launch materials state the Enterprise Edition meets ISO 27001, SOC 2, HIPAA and GDPR standards, without stating the SOC 2 type. A trust center at trust.sema4.ai exists but renders only in a browser, so no report, auditor or type has been read, and a procurement team should request the reports there. Agents run in the customer's own cloud or Snowflake account under its identity and network controls.

Several documented properties together give an auditable record. "Full data traceability and audit-ready logs" are stated on the Agent Studio page as the compliance mechanism, a record of what happened rather than a dashboard of how much. Analytics are named alongside guardrails and granular roles as the visibility and control layer of North Automations. The North product page says North connects to existing tools, data and monitoring systems, so the record can leave the platform for the customer's own observability stack. The most distinctive element is transparent reasoning chains returned alongside citations: an employee can see not only which source produced an answer but the path taken to it. That addresses why, not only what. For regulated buyers, North's whole market, the audit record is generated and kept inside their own perimeter because North deploys in their infrastructure, so there is no vendor-side log to subpoena or trust.

Sema4.ai's Actions repository documents observability out of the box: every @tool or @action run is logged and traced automatically, without instrumentation, so the trace is there when it is needed. The vendor also suggests connecting LangSmith traces with Action logs, so the action record shows what was done and the model trace shows why, in one view. Control Room provides lifecycle management, governance and audit trails, and a Work Room lets people find and supervise agents. Because agents run inside the customer's own cloud account, logs are generated within the customer's perimeter. No example trace or documentation of retention and export has been found.

North is brought to the customer's infrastructure through virtual private cloud, on-premises and air-gapped options, so Cohere itself never sees or touches customer data, and North Automations is documented as deploying securely across any on-prem or cloud environment. This is confirmed on the vendor's own security page. The engineering claim behind it is what makes this practical: North was built to run on as few as two graphics processors, described by the company's co-founder as deploying on a machine in a closet. Air-gapping is common as a promise and rare in practice, because most agent platforms assume network access for model inference; North removes that assumption by bringing the models with it. Model Vault provides dedicated deployment for the model layer specifically. This is the reason regulated buyers evaluate North at all. For a bank that cannot let data leave its firewall, most other capabilities are secondary to this one.

Agents run inside the customer's own AWS, Azure, Google Cloud or Snowflake account, with the Enterprise Edition running entirely within the customer's virtual private cloud. The Snowflake option is the distinctive one: agents act on governed data zero copy, so the data is never extracted into the vendor's environment or duplicated into a separate index. For a regulated buyer, the usual objection that grounding requires copying the corpus somewhere new does not apply. Region selection within those clouds is not documented, but for an in account deployment the customer's own region choice governs by construction.

What qualifies as ready to adopt is thin. North for Banking is the strongest item: an industry variant co-built with RBC, a real packaged offering rather than a template, but one vertical rather than a set. Agent Studio provides customizable agents and workflow automations, which is machinery for building rather than assets to take. Compass contributes pre-built connectors, which are integration components rather than agents. So a customer arriving at North gets a capable builder and, if they are a bank, a head start. They do not get a library of ready-made agents to adopt. The public surface is marketing rather than documentation, and a customer-facing agent library could exist behind the sales-led onboarding without appearing publicly. For a product sold through a six-to-sixteen-week enterprise engagement, a template catalog may simply not be public.

Sema4.ai's documentation publishes Agent templates in Studio, each a pre configured agent with its actions, a sample runbook, a setup tutorial and example conversations. Named templates include Similar Company, Search, Sales Contact Finder and Analyst, each on its own page and each doing its own job. A prebuilt actions gallery is documented separately, and the Control Room Gallery holds an organization's own agents for reuse.

The vendor says North lets enterprises "flexibly choose from Cohere's enterprise-specific generative or search models, or any external LLM". The customer chooses which model powers the product, and the choice is not limited to the vendor's own family. That is notable for a model company: it lets its agent platform run on a competitor's model, competing on the platform rather than locking the model in. Within Cohere's own range there is real breadth too: the Command family including Command A Reasoning and Command A+, North Mini Code for agentic coding, Embed for semantic search, and Rerank for retrieval precision, with the vendor documenting which product each model suits.

The customer brings its own model provider account, and the models run under the customer's contract. Sema4.ai's documentation lists three makers with recommended models: OpenAI (GPT-5.3 Codex High, GPT-5.4 High) through OpenAI or Azure OpenAI, Anthropic (Claude Opus 4.6 High) through AWS Bedrock or Azure AI Foundry, and Google (Gemini 3.1 Pro High) through the Gemini API or Vertex AI, with Snowflake Cortex models documented for the native app. Admins configure models per workspace. Accepting the customer's enterprise approved models solves a procurement problem as much as a technical one, and choosing Snowflake Cortex means inference happens where the data already lives.

Cohere names an SDK directly: enterprises interoperate with their stack using Cohere's "first-party integrations, their MCP for any custom tool, or our SDK to power bespoke agentic platforms". An SDK offered specifically for building bespoke agentic platforms on top of North is a real build surface, not merely a client library for calling models. The North product page adds that it connects to existing tools, data and monitoring systems through flexible APIs and built-in connectors, so the platform is addressable programmatically in both directions. North's MCP support uses the customer's own MCP servers for custom tools, which is North calling out to other tools. What makes North itself callable is the SDK and the APIs. Underneath sits Cohere's broader developer platform, documented as direct API access alongside North, so the engineering surface is mature rather than nominal.

The Action Server is the decisive item: an action deploys in one step as an HTTP endpoint with a public URL and Bearer token authentication, so anything outside can invoke it, with zero configuration and no infrastructure. The vendor documents connecting deployed tools to MCP clients and actions to AI applications including LangChain and OpenAI custom GPTs, with step by step guidance, so the platform is callable by an assistant the customer chose. The SDK half is equally real: Actions are written in plain Python with a managed environment, the framework is open source under the vendor's GitHub organization, and Robocorp's automation libraries come with it, so a customer can read, extend and self host the extension layer. That ease of exposure is unusual for a product whose other half is governed agents deployed in the customer's cloud.

North Automations offers a test-before-publish loop: the announcement lists testing and iterating before publishing to production, versioning to track changes over time, Plan mode to review the approach before building, per-step model choice to balance cost and performance, and analytics on usage and input and output tokens. What is missing is a readable result. No scored test set, judge, metric or comparison of versions on the same inputs is described. Cohere's red teaming and its retrieval benchmarks are run by the vendor, not by the customer, so they do not tell a buyer how their own agents perform.

Sema4.ai documents one click evaluations that test runbook changes and model upgrades, validating execution flow, actions and outputs with no coding required, created directly from successful conversations, so a real interaction becomes a regression test without anyone writing a case. The vendor calls this three dimensional validation and argues that output only testing misses logic errors. Testing model upgrades at scale against real scenarios before committing to production is comparison across versions, the question this axis asks. Studio provides the build, test and deploy loop around it, and Control Room tracks agent performance in production.

Every action path Cohere documents runs through a programmatic interface: first-party integrations to Google Drive, SharePoint, Slack, Salesforce, Gmail, Outlook and Linear, the customer's own MCP servers for custom tools, Compass connectors for retrieval, and APIs and an SDK. Nothing describes a browser, navigation, clicking, form filling or session control. The near miss is asset creation: agents producing documents, tables, charts and slideshows. Generating a file is not operating an application. North Mini Code, Cohere's agentic coding model, writes and executes code, which is also not computer use. The deployment model makes the absence coherent. A platform built to run air-gapped inside a customer's perimeter often has no external network access at all, so driving external browser sessions would cut against its main design choice. Cohere publishes no public product documentation for North, so this rests on marketing pages and announcements.

The vendor's own Actions repository states that its Robocorp automation libraries and the Python ecosystem let an agent act on anything, "from data to API to browser to desktops". Desktop automation is the clear positive case for this axis: operating an application that offers no programmatic interface means driving its interface, and browser automation sits alongside it. This is inherited capability rather than a new claim. Robocorp was an established open source Python automation company before the January 2024 acquisition, and its libraries are the platform's execution layer. Driving a desktop application through its interface is different from executing code or fetching content.